Comment 8 for bug 1859362

Using VXLAN, one can present VMs with a higher MTU than the underlying physical network, and transparently use fragmentation to fit large packets onto the network. In a sense this is virtualization of network MTU. It is slightly inefficient and can lead to unwanted side effects (e.g. in case of firewalls on the path), but there could be cases where this is useful.

The most flexible solution would be to make everything configurable by the Neutron admin:
- default MTU presented to VMs (default=1500)
- default physnet MTU (default=1500)
- minimum and maximum MTU allowed for networks (default min=68, max=9200)

Currently, the MTU for a new network is calculated based on overlay network type. For example, if the NeutronTenantMtu (neutron::plugins::ml2::path_mtu) is 1550 and you create a VXLAN network (50 bytes header), the resulting MTU for VMs is 1500. However, if you also use GRE overlay networks in the same system, VMs on a GRE overlay will get a different MTU